The work in this thesis have concluded the objectives of the research on the development and investigation of new methodologies to enable a better and reliable performance of a heartwave based biometric classification. More importantly, the proposed solutions and architectures are robust and resilient to heartwave morphological variation and heartrate variation.
Chapter 2 provided an overview of the current states-of-the-art in the domains of heartwave segmentation and extraction, and identifying discriminating features for classification. The various algorithms and methodologies are suited for different type of signals in different situations based on their limitations and constraints. In particular to the heartwave segmentation and extraction of features, none of the reported work have been attempt on heartwave signal under elevated heartrate. With regards to the classification, the review of the reported work provided an in-depth understanding of the limitations in performing heartwave based biometric in particular to heartwave signal under extreme morphological variation. This fact is the motivation of this thesis, which commits to develop new algorithms and methodologies to address extraction and classification issues caused by elevated heartrate.
In Chapter 3, a methodology consisting of Discrete Waveform Transformation integrated with heartrate dependent parameters: PR-Interval and QT-Interval has been
6-141
proposed and tested to perform 11 heartwave features identification and extraction. In total, more than 63,000 heartwaves have been extracted from 27 individuals. The dataset of the 27 individuals contains heartwave signal acquired during extreme exercise duress. Specifically, it achieved 100% detection accuracy for R-Peak and an average of 98% for the other 10 heartwave features. The use of heartrate dependent windows for PR Interval and QT Interval to support detection of wave components before and after R-Peak has achieved superior performance as compared to fixed window length. In addition, the same algorithm has been able to extract heartwave features from anomaly heartwave signal. From the database of 27 users, there are 5 individuals with anomaly heartwave signals which include signals with extended T- Peak, inverted T-Wave and irregular R-R interval. The ability to extract characteristic features with high accuracy and high sensitivity concluded the ability of the proposed solution to extract heartwave features reliably and accurately under highly variated heartrate.
In chapter 4, a novel architecture consisting of statistical based methodology of Gaussian Mixture Model integrated with Hidden Markov Model (GMM-HMM) aided with discriminating criteria of user specific thresholding score and heartrate range has been tested. The presentation of heartwave morphological variations due to heartrate has prompted a generative modelling methodology to model the joint distribution of data on individual dataset. The GMM generated model is unique to individual and the integration of Hidden Markov Model is implemented to perform discrimination of individual. Through this development, it is observed the individual matching score via loglikelihood is a linear behavior against individual heartrate. In addition, every individual has its own unique heartrate range. These understanding lead to the affirmation that individual discriminating features are not static and it vary according to its respective heartrate at the instance. To cater for the variation, the two imposing criteria, individual heartrate range and individual unique thresholding behavior are implemented. The proposed architecture achieved an accuracy of 89% from 27
6-142
individuals. More importantly, all the heartwave signals contain heartwave with extreme heartwave morphological variations. Comparing the novel proposed architecture against commonly adopted methodology of Linear Discriminant Analysis (LDA) with the same dataset, LDA achieved an accuracy of 78% with a False Positive Rate of 25%. The proposed architecture doubled the performance and achieved a False Positive Rate of 11%. This development and investigation conclude the need for proposed methodology to account for varying discriminating features. Unless an individual has no varying heartrate, methodology with fixed and static discriminating feature will more likely result in poor classification accuracy. The proposed methodology has achieved its intending objective to develop a classification model and methodology to perform biometric classification which is resilient to individual dynamic and varying heartrate.
In Chapter 5, a novel architecture categorized under neural network, has been developed and investigated. The architecture consists of an ensemble of Deep Belief Networks (DBN) connected to a module of MultiView Spectral Embedding (MSE) to combine the multiple output from ensemble-DBN into a single structure that contains significant discriminating features from multiple views. The single view structure is further input to a stacked DBN to perform classification and output via an effective and resource efficient method of Extreme Learning Machine. The performance of the proposed architecture is impressive. The proposed neural network of Deep Ensemble Architecture, is tested on 52 individuals consisting of 22 individuals with extreme heartwave morphological variations, 25 individuals whose heartwave are acquired under resting condition and 5 individuals under extreme heartwave morphological variation with anomaly heartwave signals. The proposed architecture has an accuracy of 98.9% . Equally important, the proposed architecture is capable of achieving 98.3% accuracy with a limited portion of the training data at 30%. This result is a vast improvement against most reported work where the 60% to 80% of the available data has been apportioned for training. This investigation leads to the understanding that the
6-143
increase in nodes in DBN is not necessary an advantage for highly variated signal for low signal-to-noise-ratio (SNR). DBN with lower number of nodes will results in under-fit for highly variated signals with low SNR. Conversely, DBN with high dimensional units suffers from over-fit for highly variated signals with low SNR. MSE plays a key role in identifying complementary property on views with higher order of significant. The development of the proposed neural network architecture has achieved its intended objective to develop a classification model and methodology to perform biometric classification which is resilient to individual dynamic and varying heartrate at lower proportion of training data.
6.1 Future Works
Two novel architectures, statistical and neural network methodologies, have been proposed to address heartwave based biometric constraints dues to varying heartrate and the results demonstrated promising performance. Heartwave based biometric has vast potentials to support the current and future digital age services such as IoT, cloud services and big data service. Before the solution is ready for adoption, there are works which require further exploration.